Models of higher level vision [FRG/FER]
-----------------------------------------

[GDB] Neisser's perceptual cycle again - see earlier comments.  [fg: not done]

[An understanding of higher-level vision is] reported as necessary for
continued development of models in synthetic environments  (Laird, Coulter,
Jones, Kenny, Koss, & Nielsen, 1997).

We may use Kosslyn and Koenig (1992) definition: higher-level visual
processing involves using previously stored information; lower-level visual
processing does not involve such stored information, and is driven only by
the information impinging on the retina. Thus, we will not deal here with
mechanisms used for finding edges, computing depth, etc.

Higher-level vision (HLV) is of interest for military modeling for several
reasons:

1. how LTM information will indicate incoming danger or serious change in
the environment
2. how it  will direct attention
3. how it will integrate various aspects of information, or information
occurring at different times
4. how it will be used to facilitate learning
5. how it will be used in planning and problem solving

To put it simply, HLV is at the interface between lower-level vision (LLV)
and postulated memory entities such as productions, schemata, concepts,
etc.  At the present time, this interface is poorly understood, perhaps
because LLV and LTM knowledge are not  understood in a sufficiently stable
way. (However, see Kosslyn and Koenig, 1992, for neuropsychological
hypotheses about HLV)

Most models of cognition such as Soar and ACT (actually, most architectures
reviewed by Pew & Mavor) use modeler-coded information, which avoids
dealing with the interface between LLV and LTM constructs.  Neural nets
have been used to go from pixel-like information to features or even
higher, but have not been incorporated into high-cognition models.  CAMERA
(Tabachneck-Schijf, Leonardo & Simon, 1997), and to a certain extent EPAM
(Feigenbaum & Simon, 1984; Richman & Simon, 1989), explore ways with which
features may be extracted from low-level representation, and may be
combined into LTM constructs.

This is undoubtedly an area where more research should be carried out. For
example, modeling instruction and training requires some theory on how
low-level acoustic input merges with low-level visual input and connect to
LTM knowledge.


Tabachneck-Schijf, H.J.M., Leonardo, A.M., & Simon, H.A. (1997). CaMeRa: A
computational model of multiple representations. Cognitive Science, 21,
305-350.

[Frank, you should already have the other refs.]



below the table summarizing EPAM/CHREST. I've refered to the Pew and Mavor
book more than Sloman did.

Best wishes to you and to Colleen,


-Fernand



EPAM/CHREST
-----------

ORIGINAL PURPOSE     model high-level perception, learning, and memory

SENSING AND PERCEPTION      visual, auditory
			    perceptual discrimination is in real-time
(assuming feature-based description of objects)


WORKING/SHORT-TERM MEMORY   4-7 slot STM; in some versions (e.g., EPAM-IV),
there is a more detailed implementation of a auditory (Baddeley-like) STM
and visual STM

LONG-TERM MEMORY	Discrimination net. In recent versions, nodes of
the discrimination net are used to create a semantic net and productions


MOTOR	eye movements, simple drawing behaviour

KNOWLEDGE REPRESENTATION:

DECLARATIVE  [KNOWLEDGE]  chunks, schemas  (templates); (use nodes in
discrimination net)

PROCEDURAL [KNOWLEDGE]   productions; (use nodes in discrimination net)



HIGHER-LEVEL COGNITIVE FUNCTIONS:
LEARNING	chunking, creation of schemas and of productions
		learning  is on-line (incremental) and stable against
erroneous 	data.


PLANNING	connections between templates are used in planning

DECISION MAKING		knowledge based

SITUATION ASSESSMENT	overt and inferred

MULTITASKING	serial processing; learning done in parallel

LONG-TERM MEMORY	networks of chunks, schemata and productions

RESOURCE REPRESENTATION	  limited STM capacity, limited perceptual and
motor ressources (uses time parameters)

GOAL/TASK MANAGEMENT	bottom up + 1 main goal per task simulated


MULTIPLE HUMAN MODELING		potential through multiple EPAM modules

IMPLEMENTATION PLATFORM(S)	Mac, PC (any system supporting Lisp)
				Graphical environnement supported only for MAC


LANGUAGE			Lisp

SUPPORT ENVIRONMENT		Lisp programming + editing tools
				Some graphical utilities for displaying eye
movements, structure of  discrimination tree, task
				A lot of customized code for each task modelled

VALIDATION	Extensive at many levels


COMMENTS:
EPAM models focus on a single, specific information-processing task at a
time. not yet scaled up to multitasking situations.  Used in high-knowledge
domains (e.g., chess, with about 300,000 chunks)